Gradium TTS: 81.0% Hard-Case Accuracy at 216 ms First Audio

Gradium AI shipped a new text-to-speech model on August 31, 2026 and immediately promoted it to the default across its API and Studio — no migration required, existing custom voice clones preserved. The release is notable not for incremental quality gains but for the combination of numbers: an 81.0% human-rated pass rate on a 500-sentence hard-case evaluation set and a 216 ms P50 time-to-first-audio on Coval's TTS benchmark, 170 ms faster than the model it replaces.

Modern TTS has effectively saturated on fluent prose; the failure surface has migrated to exactly the tokens a caller cannot mishear — order reference codes, email addresses with dots in unexpected positions, spelled surnames, multi-digit phone strings. Gradium's evaluation was built around precisely those payloads, which makes the accuracy gap over competitors meaningful in a way that naturalness scores are not. As we've examined before, architectural specificity now outperforms GPU scaling alone — Gradium's benchmark construction is an applied example of that principle.

Accuracy Benchmark: Design and Results

Gradium built a 500-sentence evaluation set and released it on Hugging Face under CC BY 4.0: 100 items across 10 criteria in five languages (EN, DE, FR, ES, PT). Seven atomic criteria cover spelling, acronyms, alphanumeric tokens, dates, regular numbers, large and floating numbers, and email addresses. Three composite criteria — Orders, IT Ticket, and Claims — stack several atomic criteria into a single realistic agent turn.

Scoring is human and strict: a sentence passes only when an independent native-speaker rater hears every element pronounced correctly and completely. One dropped digit fails the sentence. Audio was loudness-normalized, order randomized, and raters were capped at 40 comparisons per session with two sessions maximum and a 5-minute break enforced. Scores are pooled across all ten criteria and averaged over the five languages with equal weight. All five models were generated in August 2026 with default settings:

The Inworld row mixes two distinct model versions — TTS 2 for latency, TTS 1.5 Max for accuracy — because those are the configurations Gradium used for each respective measurement, limiting direct comparison of Inworld's joint position on the latency-accuracy plane.

Latency: Median and Variance

The 216 ms P50 figure comes from Coval's TTS benchmark across 480 runs. The operationally relevant figure is the interquartile spread: 30 ms, the tightest of the five models compared. Cartesia Sonic 3.6's 165 ms spread equals 36% of its own 454 ms median — meaning a meaningful fraction of callers wait substantially longer than the headline number suggests.

Gradium is not the lowest-latency model in the field. Inworld TTS 2 posts a 166 ms median, 50 ms faster. The claim Gradium is making is positional: the lowest hard-case failure rate at sub-250 ms first audio, with unusually low latency variance. For real-time voice agents running on Pipecat or LiveKit — both directly supported — variance matters as much as median because tail latency is what callers actually experience when the queue backs up.

Deployment and Access

The model is live with no action required from existing customers. New integrations use the Python SDK or the WebSocket TTS endpoint with existing voice IDs. The free tier provides 45,000 credits, equivalent to approximately 1 hour of TTS at 1 credit per character. On-device TTS remains a separate product line; no open weights are available for this model. The five languages covered by the published evaluation are English, German, French, Spanish, and Portuguese.

Gradium is incentivising independent validation: teams that submit complete hard-case failure reports via its Discord receive 1 million credits. The 500-sentence dataset is on Hugging Face under CC BY 4.0, so any team can re-run the evaluation against their own stack before committing — an approach that mirrors the methodology behind Chandra OCR-2's open-source benchmark.

Voice-agent infrastructure is consolidating around hard, reproducible latency and accuracy targets rather than subjective naturalness ratings. A vendor that publishes the evaluation dataset alongside the model gives buyers a concrete procurement test rather than a marketing metric — and puts pressure on competitors to match that transparency or explain why they won't.